Internal Agents Map

Agentic software factory

An agent system built around OpenAI's internal Codex. A person defines the outcome; Codex implements the change and babysits CI until green; domain-specialist agents review each change behind risk-tiered routing; and a per-change deploy agent handholds approved changes to production and builds its own monitoring dashboards.

Company
OpenAI
Approach type
Agent system
Work
Coding, Code review, CI triage, Ops
Human involvement
Human in loop
Invocation
Interactive, Background, Scheduled, Event driven
Interfaces
Desktop, Cli, Slack, Github, Skill
Deployment stage
Scaled
Evidence strength
Mixed
Entry reviewed

How it works

The workflow the sources report for this implementation.

Multiple agents, each configured as a domain specialist, review every change; the article compares this to a review by a domain expert from each relevant infrastructure team

Changes are classified by risk; high-risk changes can trigger more agent reviews or a mandated human review, while opted-in low-risk areas use an agent that auto-approves pull requests

After a human approves production, an assigned agent handholds the change to full rollout; it decides which signals mean success or failure, builds its own monitoring dashboard, and watches production signals

A perf harness sends problematic pull requests to the Synthetics A/B framework to evaluate performance implications

Agents sift through alerts and dashboards, de-duplicate signals, identify real latency regressions, root-cause them, and propose fixes

Where people stay involved

Each scope pairs its normal attention boundary with supporting evidence. See thesupervision definitions for the level mapping and limits.

  • low-risk pull request -> merge in opted-in codebase areas

    Exception only · Level 5

  • code change -> production rollout on the general path

    Work product review · Level 3

  • production alert -> proposed performance fix

    Unknown · Level unknown

Catalog interpretation

Level 5 for low-risk pull request -> merge in opted-in codebase areas; human attention boundary: exception-only.

Observed in September 2026

Catalog interpretation

Level 3 for code change -> production rollout on the general path; human attention boundary: work-product-review.

Observed in September 2026

Catalog interpretation

Unclassified for production alert -> proposed performance fix; human attention boundary: unknown.

Observed in September 2026

Implementation details

Model
unknown
Harness
Internal Codex, described as much more advanced than the external product because it is plugged into almost every OpenAI system; ChatGPT Work runs on the same harness
Sandbox
unknown
Tool access
Git repositories and GitHub, Slack, Notion, Databricks, Datadog, and internal logs and data sources
Knowledge
OpenAI moved its documentation inside the source code; internal Codex skills, some maintained by Codex itself; new engineers are directed to ask Codex during onboarding
Context management
The /goal setting lets an agent work until a goal is complete; threads run for days and spin off other agents
Credentials
N/A
Interfaces
desktop, cli, slack, github, skill

Reported results and limitations

The catalog records what the sources report, with the scope and the denominator of every figure. A qualification below limits the figure it sits under.

Headline claim

“Almost all OpenAI employees used Codex and ChatGPT Work weekly as of the September 2026 report (self-reported).”

Reported by
OpenAI
Scope
Weekly Codex and ChatGPT Work use across OpenAI employees
Method
Internal token-usage tracking by department, chart sourced to OpenAI

The source does not report the denominator of this figure.

Observed in September 2026

Key observation

Non-engineering orgs such as finance, recruitment, and legal went from about 0% to 90% Codex usage within a four-month period (self-reported)

Reported by
OpenAI
Scope
Weekly Codex use in non-engineering orgs such as finance, recruitment, and legal

The source does not report the denominator of this figure.

Observed in September 2026

Key observation

Some build-test-deploy systems saw about a 10x load increase within roughly six months (self-reported)

Reported by
OpenAI
Scope
Load on build-test-deploy pipeline systems

The source does not report the denominator of this figure.

Observed in September 2026

Lessons and interpretation

Reported opinion

Longer-running /goal tasks drove adoption; one long-running agent that spins off other agents reduces the surface a person manages

Reported opinion

Role-specific and team-specific plugins spread adoption beyond a generic coding agent

Sources and research details

Citations link to the original publisher. Each source also keeps a preserved copy in the repository, so a changed or removed page stays checkable.

  1. Inside OpenAI's agentic software factoryhttps://newsletter.pragmaticengineer.com/p/openai-software-factoryNews · Independent secondary · Last source verification: 2026-09-16
    Preserved copy
  2. The most important OpenAI announcement you probably missed at DevDay 2025https://venturebeat.com/infrastructure/the-most-important-openai-announcement-you-probably-missed-at-devday-2025News · Independent secondary · Last source verification: 2026-09-16
    Preserved copy
  3. Harness engineering: leveraging Codex in an agent-first worldhttps://openai.com/index/harness-engineering/Engineering blog · First party · Last source verification: 2026-09-16
    Preserved copy
  4. Harness engineering: Leveraging Codex in an agent-first world (discussion)https://news.ycombinator.com/item?id=48416264Hn thread · Community · Last source verification: 2026-09-16
    Preserved copy
  5. zbrock: the Codex app began as an internal prototypehttps://news.ycombinator.com/item?id=48435137Hn comment · Direct participant · Last source verification: 2026-09-16
    Preserved copy
  6. zbrock: many internal teams adopted the same practiceshttps://news.ycombinator.com/item?id=48435213Hn comment · Direct participant · Last source verification: 2026-09-16
    Preserved copy
  7. Introducing the Agents APIhttps://openai.com/index/introducing-the-agents-api/Corporate article · First party · Last source verification: 2026-09-16
    Preserved copy
  8. ChatGPT is now a partner for your most ambitious workhttps://openai.com/index/chatgpt-for-your-most-ambitious-work/Corporate article · First party · Last source verification: 2026-09-16
    Preserved copy
  9. Transcript: 'How OpenAI's Codex Team Uses Their Coding Agent'https://every.to/podcast/transcript-how-openai-s-codex-team-uses-their-coding-agentTranscript · Independent secondary · Last source verification: 2026-09-16
    Preserved copy
Research details for every claim on this page
  1. Summary
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  2. Headline claim
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Company usage claim relayed by an independent reporter; no method or denominator published. OpenAI's own 2026-09-10 launch post states nearly 100% of teams inside OpenAI, including finance and sales, use ChatGPT Work and Codex (a teams denominator, not employees).
    Reported by
    OpenAI
    Scope
    Weekly Codex and ChatGPT Work use across OpenAI employees
    Denominator
    Not reported
    Method
    Internal token-usage tracking by department, chart sourced to OpenAI
    Observation date
    2026-09
  3. Sandbox
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The article does not document an execution sandbox; unknown does not mean absent.
  4. Harness
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
    • SupportsInside OpenAI's agentic software factoryPreserved content.md, line 48 (internal Codex is a lot more advanced than its external counterpart; ChatGPT Work is powered by the Codex harness)
    • Supportszbrock: the Codex app began as an internal prototypePreserved content.md, line 12 (harness-engineering co-author: 'It was an internal prototype that looked very much like the current Codex app')
    • SupportsIntroducing the Agents APIPreserved content.md, lines 12-14 (OpenAI scaled Codex and ChatGPT for Work on the same harness; the Agents API exposes 'that same harness and infrastructure that powers Codex')
  5. Model
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The article names no underlying models for the internal harness; unknown does not mean absent.
  6. Interfaces
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  7. Tool access
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  8. Knowledge
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  9. Context management
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  10. Supporting component
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  11. Supporting component
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  12. Supporting component
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  13. Supporting component
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  14. Supporting component
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A linked participant or independent source reports the claim.
  15. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Company adoption figures relayed by an independent reporter; the four-month window carries no calendar dates. OpenAI's launch post corroborates finance-org use (month-end close reduced from days to hours) but not the 0%-to-90% trajectory.
    Reported by
    OpenAI
    Scope
    Weekly Codex use in non-engineering orgs such as finance, recruitment, and legal
    Denominator
    Not reported
    Method
    Not reported
    Observation date
    2026-09
  16. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    A VP of Engineering load estimate relayed by an independent reporter.
    Reported by
    OpenAI
    Scope
    Load on build-test-deploy pipeline systems
    Denominator
    Not reported
    Method
    Not reported
    Observation date
    2026-09
  17. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    Stated explanation by the Codex desktop lead in the reported interviews.
  18. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    Stated explanation by the Codex desktop lead in the reported interviews.
  19. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The article documents agent auto-approve for low-risk pull requests in opted-in areas and risk classification that routes high-risk changes to stricter review; the run-time exception path is not described. The harness-engineering post corroborates agent automerge of low-friction pull requests in one repository but describes no risk tiering, so the routing mechanism itself stays single-sourced.
    Observation date
    2026-09
  20. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    High
    Confidence reason
    Pipeline step 6 states that a human approves a change to go to production before the deploy agent takes over.
    Observation date
    2026-09
  21. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Unverified
    Confidence reason
    The article documents Perf Factory only up to agents that de-duplicate signals, identify real latency regressions, root-cause them, and propose fixes; whether a human reviews, applies, or lands those proposals is not stated.
    Observation date
    2026-09
Back to the directory
Know an internal agent?

Share a resource or public mention, suggest an addition, or correct an existing entry.

Suggest a resource or updateOpen a pull request